def getDeviceProps(self):
'''Get device properties'''
# write the query code
qsrc = "enum_cuda_props.cu"
qexec = qsrc + ".o"
qout = qexec + ".props"
if not os.path.exists(qout):
# check for nvcc
qcmd = 'which nvcc'
status = os.system(qcmd)
if status != 0:
g.err("%s: could not locate nvcc with '%s'" %
(self.__class__, qcmd))
try:
f = open(qsrc, 'w')
f.write(CUDA_DEVICE_QUERY_SKELET)
f.close()
except:
g.err('%s: cannot open file for writing: %s' %
(self.__class__, qsrc))
# compile the query
cmd = 'nvcc -o %s %s' % (qexec, qsrc)
status = os.system(cmd)
if status:
g.err('%s: failed to compile cuda device query code: "%s"' %
(self.__class__, cmd))
# execute the query
runcmd = './%s' % (qexec)
status = os.system(runcmd)
if status:
g.err('%s: failed to execute cuda device query code: "%s"' %
(self.__class__, runcmd))
os.remove(qsrc)
os.remove(qexec)
# read device properties
props = {}
try:
f = open(qout, 'r')
for line in f:
eline = ast.literal_eval(line)
props[eline[0]] = eline[1]
f.close()
except:
g.err('%s: cannot open query output file for reading: %s' %
(self.__class__, qout))
if props['devId'] == -2:
g.err("%s: there is no CUDA 1.0 enabled GPU on this machine" %
self.__class__)
if props['major'] < 2 and props['minor'] < 3:
g.warn(
"%s: running on compute capability less than 1.3 is not recommended, detected %s.%s."
% (self.__class__, props['major'], props['minor']))
# set the arch to the latest supported by the device
if self.tinfo is None:
bcmd = "gcc"
else:
bcmd = self.tinfo.build_cmd
if bcmd.startswith('gcc'):
bcmd = 'nvcc'
if bcmd.find('-arch') == -1:
bcmd += ' -arch=sm_' + str(props['major']) + str(props['minor'])
if self.perf_params.has_key('CFLAGS') and bcmd.find('@CFLAGS') == -1:
bcmd += ' @CFLAGS'
self.tinfo.build_cmd = bcmd
# return queried device props
return props
def getDeviceProps(self):
'''Get device properties'''
# write the query code
qsrc = "enum_cuda_props.cu"
qexec = qsrc + ".o"
qout = qexec + ".props"
if not os.path.exists(qout):
# check for nvcc
qcmd = 'which nvcc'
status = os.system(qcmd)
if status != 0:
g.err("%s: could not locate nvcc with '%s'" % (self.__class__, qcmd))
try:
f = open(qsrc, 'w')
f.write(CUDA_DEVICE_QUERY_SKELET)
f.close()
except:
g.err('%s: cannot open file for writing: %s' % (self.__class__, qsrc))
# compile the query
cmd = 'nvcc -o %s %s' % (qexec, qsrc)
status = os.system(cmd)
if status:
g.err('%s: failed to compile cuda device query code: "%s"' % (self.__class__, cmd))
# execute the query
runcmd = './%s' % (qexec)
status = os.system(runcmd)
if status:
g.err('%s: failed to execute cuda device query code: "%s"' % (self.__class__, runcmd))
os.remove(qsrc)
os.remove(qexec)
# read device properties
props = {}
try:
f = open(qout, 'r')
for line in f:
eline = ast.literal_eval(line)
props[eline[0]] = eline[1]
f.close()
except:
g.err('%s: cannot open query output file for reading: %s' % (self.__class__, qout))
if props['devId'] == -2:
g.err("%s: there is no CUDA 1.0 enabled GPU on this machine" % self.__class__)
if props['major'] < 2 and props['minor'] < 3:
g.warn("%s: running on compute capability less than 1.3 is not recommended, detected %s.%s." % (self.__class__, props['major'], props['minor']))
# set the arch to the latest supported by the device
if self.tinfo is None:
bcmd = "gcc"
else:
bcmd = self.tinfo.build_cmd
if bcmd.startswith('gcc'):
bcmd = 'nvcc'
if bcmd.find('-arch') == -1:
bcmd += ' -arch=sm_' + str(props['major']) + str(props['minor'])
if self.perf_params.has_key('CFLAGS') and bcmd.find('@CFLAGS') == -1:
bcmd += ' @CFLAGS'
self.tinfo.build_cmd = bcmd
# return queried device props
return props
def readTransfArgs(self, perf_params, transf_args):
'''Process the given transformation arguments'''
# expected argument names
PLATFORM = 'platform'
DEVICE = 'device'
WORKGROUPS = 'workGroups'
WORKITEMS = 'workItemsPerGroup'
CB = 'cacheBlocks'
STREAMCOUNT = 'streamCount'
UIF = 'unrollInner'
CLFLAGS = 'clFlags'
THREADCOUNT = 'threadCount'
BLOCKCOUNT = 'blockCount'
VECHINT = 'vecHint'
SIZEHINT = 'sizeHint'
# default argument values
platform = 0
device = 0
workGroups = None
workItemsPerGroup = None
cacheBlocks = False
streamCount = 1
unrollInner = None
clFlags = None
vecHint = 0
sizeHint = False
# iterate over all transformation arguments
errors = ''
for aname, rhs, line_no in transf_args:
# evaluate the RHS expression
try:
rhs = eval(rhs, perf_params)
except Exception, e:
g.err(
'orio.module.loop.submodule.opencl.opencl: %s: failed to evaluate the argument expression: %s\n --> %s: %s'
% (line_no, rhs, e.__class__.__name__, e))
if aname == PLATFORM:
# TODO: validate
platform = rhs
elif aname == DEVICE:
# TODO: validate
device = rhs
elif aname == WORKGROUPS:
# TODO: validate
workGroups = rhs
elif aname == WORKITEMS:
# TODO: validate
workItemsPerGroup = rhs
elif aname == CB:
# TODO: validate
cacheBlocks = rhs
elif aname == STREAMCOUNT:
# TODO: validate
streamCount = rhs
elif aname == UIF:
# TODO: validate
unrollInner = rhs
elif aname == CLFLAGS:
clFlags = rhs
elif aname == THREADCOUNT:
g.warn(
"Interpreting CUDA threadCount as OpenCL workItemsPerGroup"
)
workItemsPerGroup = rhs
elif aname == BLOCKCOUNT:
g.warn("Interpreting CUDA blockCount as OpenCL workGroups")
workGroups = rhs
elif aname == VECHINT:
vecHint = rhs
elif aname == SIZEHINT:
sizeHint = rhs
else:
g.err('%s: %s: unrecognized transformation argument: "%s"' %
(self.__class__, line_no, aname))
def readTransfArgs(self, perf_params, transf_args):
'''Process the given transformation arguments'''
# expected argument names
PLATFORM = 'platform'
DEVICE = 'device'
WORKGROUPS = 'workGroups'
WORKITEMS = 'workItemsPerGroup'
CB = 'cacheBlocks'
STREAMCOUNT = 'streamCount'
UIF = 'unrollInner'
CLFLAGS = 'clFlags'
THREADCOUNT = 'threadCount'
BLOCKCOUNT = 'blockCount'
VECHINT = 'vecHint'
SIZEHINT = 'sizeHint'
# default argument values
platform = 0
device = 0
workGroups = None
workItemsPerGroup = None
cacheBlocks = False
streamCount = 1
unrollInner = None
clFlags = None
vecHint = 0
sizeHint = False
# iterate over all transformation arguments
errors = ''
for aname, rhs, line_no in transf_args:
# evaluate the RHS expression
try:
rhs = eval(rhs, perf_params)
except Exception, e:
g.err('orio.module.loop.submodule.opencl.opencl: %s: failed to evaluate the argument expression: %s\n --> %s: %s' % (line_no, rhs,e.__class__.__name__, e))
if aname == PLATFORM:
# TODO: validate
platform = rhs
elif aname == DEVICE:
# TODO: validate
device = rhs
elif aname == WORKGROUPS:
# TODO: validate
workGroups = rhs
elif aname == WORKITEMS:
# TODO: validate
workItemsPerGroup = rhs
elif aname == CB:
# TODO: validate
cacheBlocks = rhs
elif aname == STREAMCOUNT:
# TODO: validate
streamCount = rhs
elif aname == UIF:
# TODO: validate
unrollInner = rhs
elif aname == CLFLAGS:
clFlags = rhs
elif aname == THREADCOUNT:
g.warn("Interpreting CUDA threadCount as OpenCL workItemsPerGroup")
workItemsPerGroup = rhs
elif aname == BLOCKCOUNT:
g.warn("Interpreting CUDA blockCount as OpenCL workGroups")
workGroups = rhs
elif aname == VECHINT:
vecHint = rhs
elif aname == SIZEHINT:
sizeHint = rhs
else:
g.err('%s: %s: unrecognized transformation argument: "%s"' % (self.__class__, line_no, aname))
